GPU power transient tests for UPS and generators: field procedure

How do you test GPU power transients?

Use six stages: freeze requirements, qualify input, verify measurement, run source states, judge requirements, and restore the plant. These stages follow Anvilfield procedure v1 (November 5, 2026). Add this electrical-demand test to the integrated systems test for whole-plant sequencing and cooling ride-through.

Mark the generator bus, every tested UPS rectifier input, UPS output and representative rack feeders on the installed one-line. Measure the actual bypass source, including a separate bus when fitted. Generator-terminal voltage does not establish conditions at another bypass bus.

Measure A and B demand and record PSU mode for dual-corded racks. Losing one path redistributes demand, so verify surviving-path headroom at that path rather than from aggregate hall power.

Freeze the configuration and the stop conditions

Record one-line revision, model/serial numbers, firmware, module count, governor/AVR mode, rectifier walk-in, charging limit, battery state and temperature, bypass window, synchronization settings and IT power caps. Complete generator acceptance and turnover and standalone UPS checks before integrated switching. Identify the standby generator boundary.

Write abort actions for critical-bus excursions, lost cooling, protection operation, depleted reserve and lost instrumentation. A missing limit holds the affected test. The test conductor directs the operator-approved switching sequence; the owner accepts results against the signed script.

For U.S. general-industry work, OSHA 29 CFR 1910.333(a), (b) and (c)(2), checked November 5, 2026, requires exposed live parts to be deenergized before work unless an allowed exception applies, with electrical isolation, lockout/tagout and qualified-person verification for deenergized work. Identify utility, generator, bypass and storage backfeeds before attaching or removing probes. Energized testing uses qualified personnel and the employer's applicable electrical work practices [P18].

OSHA 1910.334(c) requires qualified testers, inspected instruments/accessories, and ratings suitable for the circuit and environment; 1910.335 requires appropriate electrical protective equipment and guarding [P19, P20]. Use installed test points and a measurement chain with the required voltage, category and isolation ratings.

GPU transient versus resistive load bank: what changes?

A GPU workload adds edge speed, repetition and correlation to the sustained loading covered by conventional load-bank acceptance. Separate kW, kVA, true power factor, fundamental displacement PF with leading/lagging sign, and current distortion. Lagging reactive loading does not reproduce leading server-input behavior. A programmable bank represents only the characteristics actually delivered and measured.

Why average power and GPU nameplate are insufficient

NVIDIA describes synchronized compute/communication phases and different roles for caps, storage and power burn in its July 2025 GB300 article, updated August 27 [P03]. Its DGX H100/H200 guide lists 10.2 kW maximum system power, six PSUs and 4+2 redundancy; three powered PSUs permit reduced-performance operation [P15]. Rate feed-loss throughput separately.

NVIDIA SMI's Average Power Draw field reports the preceding second on Ampere devices (except GA100) and newer devices; its separate instantaneous field reports the last board-power reading [P04, live documentation checked November 5, 2026]. Neither measures the AC feeder. Use feeder waveforms for delivered input and telemetry for workload context.

Which standards and documents set acceptance?

IEC 62040-3:2021 edition 3 removed the nonlinear step-load type test, so a UPS performance classification alone does not establish how an installed UPS and generator will respond to repeated GPU load transients [P05].

Resolve differences without averaging incompatible limits

Apply adopted obligations, contracts and OEM constraints at their specified locations. Resolve conflicting windows and load schedules before testing; choosing the smallest percentage does not resolve different definitions.

Which workload profiles should the field procedure use?

The Anvilfield v1 library uses a 400 kW test boundary. Qualify the recipe at low approved capacity before scaling. W6, the captured deployment workload, supplies the application-specific evidence alongside AI cluster commissioning and burn-in. Record application/data hash, batch size, precision, concurrency, GPU count, scheduler, caps, software versions and smoothing.

Record wiring, PF control/sign, rack correlation and AC phase; vary approved edge phase to test unbalance. Apply OEM derating using recorded ambient, altitude, fuel and coolant/inlet conditions.

Qualify the delivered edge and plateau

For W0-W5, Anvilfield v1 requires plateau error within ±2% of target and applicable 10-to-90% duration within ±10% of its reference, including uncertainty. A full linear 1 ms ramp therefore has a 0.8 ms transition target and a 0.72 to 0.88 ms allowed interval; W2 plateaus are 100 ±2 kW and 300 ±6 kW. Declare the simultaneous three-phase power-envelope algorithm, passband, filter delay and plateau mask; use the installed three-wire or four-wire measurement method. Apply it to both reference and delivered signals; cycle-average power cannot qualify these edges.

Add expanded uncertainty to the absolute measured error and compare the sum with the allowed bound. The sum must not exceed the bound, L. A proven excess is invalid input; overlapping uncertainty is unresolved. Set the isolated-event time origin, t0, at the measured 50% crossing.

W2 averages 200 kW. W4 pairs half-cycle-offset groups to produce an ideal constant 200 kW after startup, testing correlation cancellation rather than another 300 kW aggregate peak. Preserve its startup and shutdown separately.

Approve W6’s measured peak, edge, repetition and correlation envelope before comparing runs. A workload-file hash alone cannot establish identical electrical demand.

Keep repetition and thermal conditions reproducible

Restore electrical, battery and thermal conditions between repeats. Extend capture through recovery/recharge with 1 s environmental trends. Three repeats expose variation, not statistical reliability.

How do you capture synchronized UPS and generator waveforms?

Anvilfield v1 starts its 1 ms program at 100 kS/s per simultaneous voltage/current channel and 20 kHz usable measurement-chain bandwidth with documented anti-alias filtering. Revise acquisition for faster edges; this millisecond recipe is not an IEC minimum or an emissions test. EPMS and power metering provide operational context.

Prove timing before interpreting causality

Verify common-clock or supported synchronized acquisition with an isolated marker before/after each run. Report measured skew and sensor delay. Ordering within combined timing/filter uncertainty is unresolved; UTC does not remove controller scan delay.

Do not confuse instrument specifications

HIOKI's PW8001E15-5XB brochure separates ADC sampling, numerical updates and optional D/A waveform output [P14]. Qualify sensor, analyzer, output and logger together, including rented synchronization options and firmware.

Anvilfield estimates 0.96 GB for 24 waveform signals over 100 s at 100 kS/s and four bytes per value, excluding timestamps and copies. Method: four three-phase locations require twelve voltage/current pairs. Multiply the number of signals by capture seconds, samples per second and bytes per value: 24 × 100 × 100,000 × 4 bytes.

Add bypass, A/B, DC and marker channels to the acquisition plan. Rehearse gap-free capture and readable export; segmented or compressed data cannot establish uncaptured fast extrema.

Which source states and transfers must you test?

Qualify acquisition on utility first; compare identical input hashes and starting conditions across states. In parallel plants, record each set’s kW/kVAR and sharing mode using the generator paralleling and load-sharing procedure; aggregate bus power hides unequal loading.

Capture pickup, recharge, and retransfer separately

Mark source loss, discharge, generator availability, ATS state, rectifier walk-in and charging separately. Include conversion losses, recharge and non-IT generator loads. Follow the full return sequence and controller timers. For gas engines, rotary UPS or DC UPS, replace limits and state definitions with the installed architecture; IEC 62040-3:2021 excludes rotary and DC UPS from its scope [P05].

Bypass and degraded capacity are separate scenarios

Use UPS and STS hold points for approved transfer acceptance. Verify phase/frequency tracking and bypass permissive. Precheck remaining capability before removing a generator, module or path. Vertiv SL-80617 recommends double conversion and avoiding maintenance bypass directly onto generators with high AI fractions [P17]; the installed operating procedure still defines authorized switching.

What is a pass, hold, or stop?

Anvilfield's GPU transient procedure v1 judges critical-load continuity, generator response and bypass synchronization separately; a powered load alone does not prove synchronized-transfer availability. Pass requires every mandatory criterion to be met. A recorded breach fails; deficient input/data invalidates evidence.

Use hold for a missing prerequisite or limit. Not applicable needs a topology reason; omitted applicable cases remain unproven. Stop identifies the operating action.

OSHA 1910.334(b)(2), checked November 5, 2026, prohibits repetitive manual reclosing after protective operation and requires determining that the circuit can be safely reenergized [P19]. A protection trip stops the scenario for that determination; it is not a request to reset and repeat.

Put every numerical limit on the signed script

Sign the location, units, reference, window, excursion duration, recovery band/deadline, authority/revision, decision owner and stop action. Select project bands before testing. Anvilfield estimates a project-selected ±5% band at 415 V as 394.25 to 435.75 V and a project-selected ±1% band at 50 Hz as 49.5 to 50.5 Hz.

Recovery begins at band re-entry sustained for the declared dwell. Use the table's analysis dwell unless the contract specifies another. For W2, judge the burst envelope and persistent oscillation, then settling after its final edge. Its 25 Hz input has no quiet second between steps.

Treat uncertainty straddling a limit as unresolved. Retain the uncertainty with the recovery result.

Worked example: what does the published 2.4 MW test show?

Vertiv SL-80617 (07/26) supplies the worked example’s inputs and results. Calculations use its load-side boundary.

Method: Anvilfield estimates total nominal IT power by adding AI and other IT power. Divide AI power by total IT power or generator nameplate to find each AI fraction. Multiply nominal AI power by the 1.5 peak factor, then add other IT power to find peak total IT power.

Subtract nominal IT power from peak IT power to find the increment, then divide by nominal IT power for the percentage increase. Divide peak IT power by generator nameplate for the nameplate comparison. Divide real power by the assumed 0.95 leading power factor to find apparent power: S = P / PF. The load-side boundary includes IT demand; account separately for losses, charging and non-IT demand when calculating generator output.

The denominator changes the meaning of “75% AI”

Vertiv's approximate 75% synchronization threshold uses total IT load as its denominator. The scenario table also gives AI fractions of generator rating. Use the denominator shown in each row; the threshold is specific to this test system.

Use the detailed result, not the simplified summary

Vertiv's September 2026 blog combines voltage and frequency in an approximate ±0.5% summary; the full paper separates them as shown above [P01, P17]. Its worst-case bypass difficulty despite small frequency variation shows why phase tracking needs separate evidence.

Where should smoothing act: GPU, rack, or UPS?

Locate each control before comparing results. GPU controls change demand; rack energy storage and ride-through buffers the rack boundary; UPS storage changes upstream input. The measurement boundary determines which improvement is visible.

Freeze controls and change one variable at a time

Export the supported controls and units, recording unavailable fields rather than substituting zero. NVIDIA DCGM smoothing fields require the documented access privileges [P12]. For UPS comparisons, match downstream demand; for GPU or rack controls, repeat the same workload request and measure the changed feeder profile. Hold source state and battery starting condition fixed.

Judge throughput alongside power using power capping and oversubscription criteria. Record workload progress for each compared setting.

Size the power exchange and repeated energy separately

Anvilfield estimates that flattening W2 at 200 kW requires up to 100 kW discharge and 100 kW absorption, with 1.95 kJ discharged per cycle and 0.81 kWh discharge throughput over 1,500 cycles. Method: integrate 100 kW over the 19 ms high dwell plus two positive half-ramp triangles, each 0.5 ms wide and 100 kW high; multiply by cycles and divide kJ by 3,600 for kWh. Equal ideal charging yields zero net energy change.

Measure DC pulse-current capability and assess battery life from the installed storage duty. Measure storage current, temperature and energy in both directions, then quantify recharge peak and restoration time. Conversion losses, reserve, control behavior and OEM limits determine actual storage duty.

What must the synchronized recovery plots show?

Align fast-event, burst and transfer views. This chart compares reported extrema; use Vertiv's source screenshots for the recorded traces.

Fast view: use -10 to +50 ms around the measured W1 edge, with per-phase raw V/I, instantaneous three-phase p(t), marker and trigger. Overlay the commanded ramp and qualified envelope separately, marking 10/50/90% crossings. Annotate bandwidth, sampling, probe delay, skew and clipping. Anvilfield estimates 6,000 sample intervals across that window at 100 kS/s.

Recovery view: start at -1 to +10 s and extend until recovery is established. Align feeder kW, per-phase RMS voltage, frequency, DC current and bypass permissive/phase error. Draw approved bands and mark extrema, first and sustained re-entry. Declare RMS/frequency algorithms, windows and latency.

Flag frequency as invalid when voltage loss or distortion defeats the estimator. Never plot a held last value as continued synchronization.

Burst view: show -10 to +90 s for W2, distinguishing burst onset from its final edge. Preserve extrema in long views and label continuous raw, segmented or compressed coverage. Extend the separate transfer view through pickup, walk-in, recharge and retransfer; mark source availability, ATS contacts, UPS mode and smoothing state.

Separate likely causes with a discriminating trace

Source-frequency oscillation with stable UPS output points upstream; compare rectifier demand and bypass state before blaming governor or AVR alone. Good source voltage with a feeder excursion requires UPS-output comparison. A smaller-than-commanded step with job slowdown requires cap, thermal and application logs. Immunity to an input never delivered remains unproven.

When must you retest?

Repeat affected profile/state combinations after relevant workload, control, storage or topology changes. Anvilfield v1 proposes an evidence review every 12 months, not mandatory annual GPU testing. Preserve failed and invalid runs with corrective-action IDs; one attractive replacement waveform does not close a finding.

Close a finding with evidence

Close findings after required results pass, exclusions are explicit, as-left settings match and automatic operation is restored. Retain failed originals; missing samples require reacquisition.

What should the scope, schedule, and estimate include?

Anvilfield estimates $24,000 for this two-shift support package or $31,050 with one retest shift, before excluded site costs.

Method: multiply labor hours by the hourly rate, then add instrument and travel allowances. Anvilfield selected these estimating inputs on November 5, 2026; vendor quotations must be obtained separately.

The support scope excludes load equipment, temporary distribution, fuel, switching crew, OEM attendance, taxes, premiums, outage cost and owner fees. Allow for access and battery restoration.

Find and qualify the commissioning work

An AI hall addition, GPU-generation change, altered generator topology or revised power-control strategy defines a commissioning scope trigger. Commissioning teams planning generator service territories use PorchRocket Asset Search for reviewed lists of visible generator sites and their owners of record. Operating drawings, baseline records and GPU rack power and cooling readiness define the technical work package.

Build the schedule by multiplying the number of profiles by repetitions and compared settings within each source state. Add the captures across states, then add transitions. Anvilfield estimates W2/W3/W4 across U and G, three repeats and two settings at 36 captures. At 100 s each, recorder time alone is one hour.

Add other profiles, switching, battery restoration, setup and reruns to establish the full schedule. More measurement boundaries, permitted switching windows and recovery time determine schedule and cost.

Field checklist and turnover dataset

Check each state change and turnover.

• Confirm electrical work controls, qualified testers and inspected/rated equipment; missing controls hold setup.
• Match the one-line, boundary and prerequisite generator/UPS records; mismatches hold the scenario.
• Name each limit’s source, revision, units, location and window; undefined criterion holds acceptance.
• Hash workload/replay inputs; changed input gets a new run ID.
• Qualify measured input fidelity; wrong profile invalidates the run.
• Measure both affected rack feeds and actual bypass source; unmeasured paths remain unproven.
• Rehearse channel count, duration, export and gap-free capture; inadequate coverage requires correction.
• Verify sensor IDs, ratios, orientation, calibration and marker alignment; clipping or failed timing requires reacquisition.
• Capture UPS mode, charging, walk-in, bypass and smoothing settings; undocumented changes restart comparison.
• Record generator count, sharing and environmental state; capability changes require script review.
• Verify battery starting condition and reserve; inadequate reserve defers testing.
• Record cooling, workload progress and errors; investigate throttling and service failure separately.
• Record output, generator and bypass verdicts for each state; unrun applicable states remain unproven.
• Retain failed files and issue IDs; repeat corrected cases and affected transitions.
• Match final settings and automatic state to turnover record; mismatch holds release.

Reusable input CSV and readme

The CSV defines W0-W5: each row starts a linear ramp to command_kw over ramp_ms, then holds. W4 offsets are already in time_s; wrap rows modulo 40 ms for the steady pattern after approved startup. Revise file/hash for extended W1/W5 dwells. W6 retains its acquired waveform.

Download W0-W5 command CSV and input readme. Terminal rows mark endpoints; the next cycle takes precedence. W2/W3 start at 100 kW. W4's startup begins with four 25 kW groups, then establishes the cyclic pattern before steady capture.

profile_id,time_s,group_id,command_kw,ramp_ms,repeat_count
W0,0,all,100,0,1
W0,60,all,100,0,1
W1,0,all,100,0,3
W1,10,all,300,1,3
W1,20.001,all,100,1,3
W1,30.002,all,100,0,3
W2,0,all,300,1,1500
W2,0.02,all,100,1,1500
W2,0.04,all,100,0,1500
W3,0,all,300,1,60
W3,0.5,all,100,1,60
W3,1,all,100,0,60
W5,0,all,100,0,3
W5,10,all,300,10000,3
W5,30,all,100,10000,3
W5,50,all,100,0,3
W4,0.000,1,75,1,1500
W4,0.020,1,25,1,1500
W4,0.040,1,25,0,1500
W4,0.010,2,75,1,1500
W4,0.030,2,25,1,1500
W4,0.050,2,25,0,1500
W4,0.020,3,75,1,1500
W4,0.040,3,25,1,1500
W4,0.060,3,25,0,1500
W4,0.030,4,75,1,1500
W4,0.050,4,25,1,1500
W4,0.070,4,25,0,1500

Preserve a reconstructable record

Store run_id, profile_id, state_id, repeat, t_utc, t_rel_s, channel, location, phase, quantity, unit, value, sample_interval_s, window_s and quality_flag. Metadata include nominal voltage/frequency, ratios, signs, timing uncertainty/method, marker origin, delay corrections, gaps, source availability and bypass state.

Keep the limits register with criterion_id, source_document, revision, state_id, location, quantity, window_s, limit_curve, excursion_duration_s, recovery_band, recovery_deadline_s, uncertainty, decision_owner and stop_action. Preserve its approved version with every run. Retain hashed binaries, CSV exports, configurations, scripts, logs, calibration, time checks, defects and as-left settings; version corrections. Define RMS/frequency/settling calculations in the readme and distinguish observed, manufacturer-reported and Anvilfield-calculated data.

How to cite this page: Anvilfield, GPU power transient tests for UPS and generators: field procedure, 2026, https://anvilfield.com/field-guides/datacenter/gpu-power-transient-test-plan/.

Related tools

FAQ

People also ask

• What is an integrated systems test?
• How long is a generator load bank test?
• What is an NFPA 110 acceptance test?
• What is a commissioning hold point on a UPS?

Sources checked

• P01: Vertiv, Powering AI at scale September 3, 2026 blog; benchmark, four scenarios, and discrepancy with full paper. Accessed 2026-11-05.
• P02: Vertiv, Pushing data center limits July 30, 2026 landing article; 2.4 MW dual-UPS study and access to detailed paper. Accessed 2026-11-05.
• P03: NVIDIA, How new GB300 NVL72 features provide steady power for AI July 28, 2025, updated August 27, 2025; correlated loads, cap/burn/storage phases, 65 J/GPU, reported 30% peak reduction in a GB200 rack fitted with the new energy-storage-enhanced shelves. Accessed 2026-11-05.
• P04: NVIDIA System Management Interface documentation Live documentation snapshot; previous-second average board power, power-smoothing controls and units. Accessed 2026-11-05.
• P05: IEC 62040-3:2021 Edition 3.0, April 21, 2021; published UPS test scope and edition changes. Accessed 2026-11-05.
• P06: IEC 61000-4-30:2025 Edition 4.0, October 24, 2025, corrected July 2026; PQ measurement scope/classes and transducer qualification boundary. Accessed 2026-11-05.
• P08: Cummins, Transient Performance of Generating Sets Bulletin 5544386 Rev. 2/25; source/load interaction, recovery definitions, BMEP steps, ISO 2018 interpretation. Accessed 2026-11-05.
• P09: Caterpillar, Transient Performance Specifications for Diesel Generator Sets LEXE22885-00, September 2024; cites ISO 8528-5:2022, custom step/G4 distinction, inertia/configuration effects. Accessed 2026-11-05.
• P12: NVIDIA DCGM field identifiers Live snapshot; field IDs for smoothing, mW/s ramp units, ms hysteresis, access privilege requirements. Accessed 2026-11-05.
• P13: BSI, BS ISO 8528-5 current edition record Current 2025 release, May 8, 2025; edition history and replacement of 2022. Accessed 2026-11-05.
• P14: HIOKI PW8001 brochure PW8001E15-5XB, information correct October 3, 2025; module rates/bandwidth, D/A waveform feed versus numerical update and saved-file distinctions, 5 Mword memory/compression, synchronization options and firmware qualifications. Accessed 2026-11-05.
• P15: NVIDIA DGX H100/H200 user guide, introduction Live user-guide snapshot; 10.2 kW maximum system rating, six PSUs and 4+2 redundancy, input supply specifications. Accessed 2026-11-05.
• P17: Vertiv, Adapting data center power infrastructure to AI workloads, full white paper SL-80617 (07/26), 14 pages; actual test setup, Scenario 2 inputs/results, synchronization difficulty, and model-specific recommendations. Accessed 2026-11-05.
• P18: OSHA, Selection and use of work practices 29 CFR 1910.333(a), (b), (c)(2); general-industry deenergizing exceptions, multiple energy sources, lockout/tagout, verification and qualified energized work. Current regulatory text checked 2026-11-05 through the official OSHA page.
• P19: OSHA, Use of equipment 29 CFR 1910.334(b)(2), (c); safe reenergization after protective operation, no repetitive manual reclosing, qualified testers, inspections and equipment ratings. Current regulatory text checked 2026-11-05 through the official OSHA page.
• P20: OSHA, Safeguards for personnel protection 29 CFR 1910.335(a), (b); task-appropriate electrical protective equipment and guarding/barricades. Current regulatory text checked 2026-11-05 through the official OSHA page.